好的，我需要帮用户总结这篇文章的内容，控制在100字以内。首先，文章主要讨论了AI系统中的一个问题，即决策执行的权限模糊性。作者提出了一个叫做Authority Encoding Risk（AER）的新概念，用来衡量这种模糊性带来的风险。 接下来，文章指出当前的AI系统虽然能评估决策的正确性、政策符合性和安全性，但无法可靠判断决策是否符合实际的权限约束。这种差距存在于自动化系统、AI辅助决策、机构工作流程以及保险和损失建模等多个领域。 最后，作者提到这篇论文详细分析了权限模糊如何转化为风险，现有框架为何无法捕捉到这一点，以及如何在损失发生前进行测量。他们认为这是一个值得重视的新风险类别，并计划进一步完善模型和收集更多证据。 总结一下，文章的核心内容是引入AER概念，指出当前AI系统忽视了权限模糊带来的风险，并强调这一问题的重要性。 </think> 文章提出Authority Encoding Risk（AER）概念，探讨AI系统中决策执行权限模糊带来的风险。现有系统无法可靠判断决策是否符合实际权限约束，导致潜在风险在多个领域中存在。文章分析了权限模糊如何转化为风险，并强调这一新风险类别的重要性。 2026-4-1 08:41:6 Author: www.reddit.com(查看原文) 阅读量:2 收藏
Most AI discussions focus on correctness.
Accuracy. Alignment. Output quality.
But there’s a more fundamental problem underneath all of that:
Who — or what — is actually allowed to execute a decision?
---
I just published a paper introducing:
Authority Encoding Risk (AER)
A measurable variable for something most systems don’t track at all:
Authority ambiguity at the moment of execution.
---
Today’s systems can tell you:
• if something is likely correct
• if it follows policy
• if it appears safe
But they cannot reliably answer:
Is this decision admissible under real-world authority constraints?
---
That gap shows up in:
• automation systems
• AI-assisted decisions
• institutional workflows
• underwriting and loss modeling
And right now, it’s largely invisible.
---
The paper breaks down:
• how authority ambiguity propagates into risk
• why existing frameworks fail to capture it
• how it can be measured before loss occurs
---
If you’re working anywhere near AI, risk, infrastructure, or decision systems — this is a layer worth paying attention to.
---
There’s a category of risk most AI systems don’t even know exists.
This paper represents an initial formulation.
Ongoing work is focused on tightening definitions, expanding evidence, and strengthening the model.
如有侵权请联系:admin#unsafe.sh